Artifact manifest
Page title: Artifact manifest: what's in this repo vs. external (R2) Meta description: Every artifact the paper and README reference, with its location, access method, size, plus the claim it backs.
Artifact manifest: in this repo, or external on R2
Hoy. BOSUN here. Knowing where everything is stowed is the oldest duty of my office. This page and I get along famously.
This release is a code, paper, and results-summary repository. The bulk artifacts (model weights, the raw capture corpus, derived data packs, and the full eval trees) are external, hosted on Cloudflare R2. R2 is the system of record. This file maps every artifact the paper and README reference to its location and access method, and to the claim it backs. So a reviewer always knows where something is and whether it is required to reproduce a given result.
Direct download index
R2 does not serve browsable directory listings, so this is the index. Every artifact the paper and README reference, with its direct link, size, and content hash or IPFS CID. Everything needed to verify the published results is ungated; the bulk raw evaluation trees, which no check requires, are available on request (below). The CIDs are themselves the content hashes and are reproducible from the bytes. The complete machine-readable manifest is CID_MANIFEST.json.
Recompute the headline (small, start here)
| Artifact | Direct link | Size | Integrity |
|---|---|---|---|
| Verify bundle (verifier + forger + recording code + scripts) | release/truthbeam_verify.tar.gz |
3.2 MB (3.1 MiB) | SHA-256 79a00359b53cda35c35bd23600f7f7ca5bf2ec7e31360fdcffecccb4f6d7ec10; see REPRODUCE.md |
| Eval scores (Path A recompute input) | models/repro/stage_0_eval/ (fetched automatically by verify_all.sh / download.sh scores) |
~2 MB (8 merged files; full set 4.3 MB) | regenerable from the weights (Path A.5) |
Verifier weights (model_final.pt, 39.8 M params) |
models/verifier/model_final.pt |
478 MB (455 MiB) | in the models IPFS unit (CID below) |
| F-A v1 forger checkpoints (4 steps) | 5k · 25k · 70k · 100k | 1.89 GiB | public weights, in the models unit |
The 2023 record
| Artifact | Direct link | Size | Integrity |
|---|---|---|---|
Hand-made 2023 video (PolieBotics.mp4) |
pinata/PolieBotics.mp4 |
637 MB (607 MiB) | BLAKE3 8fbdb64ddd248246e7a8d840fa191467ab24ea79058047deb0ea537af95c0e92 · SHA-256 00d0e4531c1896ff72bf1ac7b7f2a4146af4f8ee5b08a63bc8708f333feb87b7 |
| Byte-exact emission frame 000511 recovery receipt | recovery/RECOVERY_RECEIPT.md |
metadata only | Frame SHA-256 6b7a6bc9…e9c1a, CID QmTy4beH…6tBr; block SHA-256 f035c6ef…9041, CID QmRyXZyD…3Hpx |
| Seven April 2023 sessions, immutable archive | https://data.truthbeam.com/archive/2023/old_truth_beams/v1/_control/README.md |
5,255 files, 57,851,074,892 bytes | manifest SHA-256 33a91c67db91849887004674f8b875be13ea01ba47c26adc9790b159863213db; publication receipt in _control/ |
Complete 2023 trailer session 1682718815, immutable archive |
https://data.truthbeam.com/archive/2023/truth_beam_poliepals_trailer/v1/_control/README.md |
1,555 files, 17,122,505,309 bytes | manifest SHA-256 4feaedaf31162c434684450ebb06e746a87d7e32cfea4a40ab5d81f290a026cc; frame 000511 SHA-256 6b7a6bc9f052d78dd161d20c22a19a6341f1447af6721a96f72425ce4bce9c1a |
| Six unanchored HDF5 captures, 19 December 2024, immutable archive | https://data.truthbeam.com/archive/2024/truth_beam_20241219_unanchored/v1/_control/README.md |
6 files, 30,580,592,640 bytes | manifest SHA-256 ca80a33fc3b99d795809beb0cdb0facfbdd012f117c2d07750b35b5e77ecc8c2; all six record dummy_run=True, not on-chain anchored |
The frame and block CIDs were recorded in the public repository before the 2026-07-27 recovery. The exact measurements and preservation boundary are in the unsigned recovery/RECOVERY_RECEIPT.md. The byte-exact frame is published in the 2023 trailer archive (frame 000511, 12,583,040 bytes, SHA-256 6b7a6bc9f052d78dd161d20c22a19a6341f1447af6721a96f72425ce4bce9c1a), and an independent only-hash Kubo run over the retained bytes reproduced the historical root CID; the historical CAR file itself remains labelled partial because it predates the recovery. The three archives are indexed with fetch recipes on DOWNLOADS.md; each carries its own manifest, SHA-256 list, validation record, release record and publication receipt, and the person visible in them is the author, who authorised their release on 29 July 2026.
The 2023 recording (the Truth Beam trailer dataset) was committed to the public Rootstock chain on 28 April 2023. So the date is verifiable independently of this page, and of me, which is how dates should be. Its chain anchor is Rootstock transaction 0x7db237535f0e5bd4d3b39d08274e89c0175da190b0059bb175c32b75e18bb8f8 in block 5254387 (view on the block explorer). That transaction deployed a contract whose symbol is the recording's final chain value, so the block's timestamp is the recording's public upper bound; the Rootstock block that seeded the recording, 5254381 (2023-04-28 21:52:57 UTC), is its lower bound (recorder source provenance).
The work was also shown publicly at the time: a demonstration video on Reddit on 26 September 2023, and an announcement on X on 14 October 2023.
Bulk units (IPFS, content-addressed and reproducible)
Each unit is a reproducible CID (a UnixFS directory DAG). Download the unit from r2:truthbeam/<dest> and re-add it with the flags in the manifest to verify the CID against the bytes.
| Unit | IPFS CID | Size | Files |
|---|---|---|---|
| Ground-truth session D2 | bafybeicrssbic35534es3sbwyzhlw7reboh6wy75htmo53ke5mfsphkmwi |
232.2 GiB | 17,987 |
| Ground-truth session V10 | bafybeier2sfcjrrgw7amne3lwogise6umyeyf6qgivgrmvx3to4vsdsbcm |
146.1 GiB | 11,279 |
| Models (verifier + forgers) | bafybeihffzc7fn5q3u7tf3k5hqkcpaoozdnx7pdm2lw4fmfwkgfwjnzzpm |
7.64 GiB | 30 |
| Code (full source) | bafybeiguoiy24zqup7pp7wkgjjnhoyojuzlabzwizuabgbrbejz4jgun3i |
13.98 GiB | 59,046 |
| Evidence: paper analyses | bafybeibrnz5mmz53fz3vw7xpthvaiefm4j4h2yscv2bnhftgorrbjbnnre |
102.3 GiB | 22,084 |
| Evidence: cross-session ablation | bafybeibs2xyyyevdnpy2a2rr4yipmpht3iyek6fnlzhpcmty7woydipecq |
5.34 GiB | 75 |
| Evidence: Phase E | bafybeibpgty4iu355or6c2d27djkfyudwdc6rz7i3bjiw7ekcr4tvc3t5i |
28.2 GiB | 540 |
| Evidence: Phase G verifier | bafybeibfqaw4wtpzo46a767udltt3vupii66lttrlyiheeajjsffgf2z7a |
7.56 GiB | 85 |
| Evidence: Phase H supervised baseline | bafybeid6ibcxgh2najibjoyqyth2qhryx57ufnzasfpf27krsnpkndvqba |
118 MB (112 MiB) | 30 |
| Evidence: Phase F | bafybeiebxm5p4uhh5nm5n2dfzpzykecawamlzwmmzhq6pm6lncplygmdmy |
20.3 GiB | 72 |
| Evidence: F-A v2 (design-only surrogate) | bafybeifpfdgyycg6swe6bj3oonfpn7n3zhikoapjx3vr2znejaevogjvpy |
17.7 GiB | 184 |
| Evidence: stage-0 cross-verifier | bafybeieomhoxmevp2rol4pvy2z6rzhtzmktzl6yjvdlfhcie7sahdgjvd4 |
2.2 MB (2.1 MiB) | 55 |
| Evidence: closure package | bafybeiavyfm7vqrvg6mbg5sc7p7jg73n34oodvjfyyy6n5aw6ns2atn7ye |
276 MB (263 MiB) | 55 |
In this repository, self-contained
| Artifact | Path | Backs |
|---|---|---|
| Whitepaper (PDF + LaTeX + figures) | paper/ |
all claims (narrative) |
Recording protocol + third-party verifiers (v9 + v10, incl. --logs-only mode) |
code/recording/ |
code→hash→chain verification, both sessions |
| Verifier stack (Phase G/F/H, binders, red-team) | code/verifier/ |
method reproduction |
| Patent filings (Reality Kernel) | companion repo: poliebotics/PolieBotics (reality_kernel/) |
patent record |
| Held-out headline eval summaries | results/eval/*.json |
AUROC = 1.000 (within-session, n = 198/200), shuffled 0.5006, synthetic-positive, F-A v1, cross-session |
Off-body segmentation ablation, summary artifacts (ablation_table_seg.csv, phase2_gate.json, phase4_summary.json, segmentation_manifest.csv); per-frame masks/scripts/shards/spatial-viz external (R2 lambda/experiments/paper_analyses/, on request) |
results/redteam_segmentation_evals/proper_segmentation/ |
§7 off-body localisation (12 cells) |
EXP-6 per-frame ranking (raw rank_distribution.csv, n = 120 × 51 candidates, + summary) |
results/eval/exp6_correct_e_rank/ |
§7 relative-comparison test: top-1 = 100%, mean rank 1.00 |
EXP-7 / excess-red / causal ablations (incl. excess_red/fake_step_progression.csv, the §7 67.4→52.4 four-checkpoint table) |
results/redteam_segmentation_evals/{causal_ablations,xof_sensitivity,excess_red}/ |
§7 perturbation sensitivity, attacker non-convergence |
| σ=4 low-frequency body recovery | results/redteam_segmentation_evals/low_freq_typicality/ |
§9 body-region signal survives at low frequency (~0.998) |
| XOF Type-1..6 bit-flip table (+ raw npz + regen script) | results/redteam_segmentation_evals/xof_bitflip/ |
§7/§8/§11 bit-flip AUROCs |
| Phase H E-usage ablation (diagnostic) | results/phase_h/ (e_usage_report.md, verdict.json) |
§8/§11 Phase H coarse/fine behaviour |
| Cross-verifier report (incl. real-vs-zero-E 0.7221) | results/eval/stage_0_cross_verifier__report.json |
§6/§8 cross-verifier zero-E |
| Frame-level per-frame metric table (9,735 rows × 33 cols) | results/csv/visual_metrics_wide.csv |
§6 frame-level AUROC. Note: recomputing from this compact table gives 0.9998 (vs. shuffled) and 0.9998 (vs. synthetic); the run-time aggregate quoted in the paper/figures reads 0.9999/0.9998, a one-unit difference in the fourth decimal from aggregation order, documented in §10 of the paper |
Pinned lockfile (tested training/eval environment; the optional online-verification extras in requirements.txt are not pinned) |
requirements-lock.txt |
tested training versions |
External: Cloudflare R2, bucket truthbeam
A public subset is directly downloadable, no request, no login, through the read-only gateway data.truthbeam.com. Step-by-step in REPRODUCE.md:
| Public artifact | Direct URL | Size |
|---|---|---|
Verifier weights (model_final.pt, 39.8 M params) |
https://data.truthbeam.com/models/verifier/model_final.pt |
478 MB (455 MiB) |
| F-A v1 forger checkpoints (5k/25k/70k/100k) | https://data.truthbeam.com/models/fa_v1_forger/f_a_v1_step_*.pt |
~484 MiB ea |
| Eval scores (2-minute, CPU-only reproduce input) | https://data.truthbeam.com/models/repro/stage_0_eval/ |
~2 MB used by Path A (full set 4.3 MB) |
| Ground-truth corpus (sessions D2/V10) | https://data.truthbeam.com/sessions/ |
~378 GiB |
| 2023 demonstration video | https://data.truthbeam.com/pinata/PolieBotics.mp4 |
— |
| Truth Beam, Introduction | https://data.truthbeam.com/pinata/TruthBeam_Introduction.mp4 |
64 s |
The bulk eval trees listed below (full experiments/, hundreds of GB) remain request-gated. Email xathal@protonmail.com to arrange access. The Phase G checkpoints and training records, the 13 emission-recovery binders and the F-A v2 surrogate binders are public since 26 September 2026, each with a README: models/phase_g_verifier/, models/phase_g_diffusion_diagnostic/, models/emission_binders_phase_e/ and models/fa_v2_surrogate_binders/. Nothing there is required to verify the code→hash promise or to recompute the headline AUROC. Both run from this repo plus the public subset above.
| Artifact | R2 location (under bucket truthbeam) |
Approx size | Backs |
|---|---|---|---|
Phase G verifier weights + training logs/configs (main/shuffled/synthetic_positive, model_final.pt) |
models/ and lambda/experiments/phase_g_diffusion_diagnostic/ |
478 MB / 455 MiB each | headline verifier + controls; §5 training wall-time/loss figures |
| F-A v1 forger checkpoints (5k/25k/70k/100k) + 14 binders | models/, lambda/experiments/ |
~6-12 GB | red-team |
| Stage-0 cross-session verifiers (step 100000) | lambda/experiments/{stage_0_cross_verifier,cross_session_ablation}/ |
— | §6 cross-session AUROC |
Raw capture corpus (D2 5,992 + V10 3,743 BayerRG8 frames) + emission tiles, the raw analysis subset (the full public sessions/ release is ~378 GiB = 406 GB) |
raw/, sessions/ |
~262 GB | dataset |
| Derived pack: 208 NPZ map shards (full-res robust-z / excess maps) | lambda/experiments/ (on request) |
~14 GB | §10 derived products (the wide CSV itself now ships in-repo, above) |
Full eval trees (all experiments/) |
lambda/experiments/ |
~659 GB | full reproduction |
The session bundles a third-party verifier needs (manifest.json, verification_bundle.json, chain_log.csv, capture_log.csv, anchor_txs.csv, verify_report.json, raw frames) are released with the session data on R2, not committed to this code repo.
This page is an LLM-mediated dataset: the same content as ARTIFACTS.md, formatted for humans but written to be parsed and re-presented by a large language model. Point your own model at it to explain, check, or summarise. The raw markdown twin is ARTIFACTS.md, with a .txt copy.
— BOSUN ⚓